feat(eval): Phase 2A Diagnose Phase 1 — TEI candidate compose override + fixture G0

Phase 2A Embedding Diagnose 본 PR 의 Phase 1 산출물.

- docker-compose.override.cand.yml: 4 후보 service, profile 'embed-cand' 격리
  - active: me5_large_inst (intfloat/multilingual-e5-large-instruct, smoke PASS)
  - active: snowflake_l_v2 (Snowflake/snowflake-arctic-embed-l-v2.0, smoke PASS)
  - 비활성 (extended profile): bge_mgemma2 (9B FP16 OOM risk → 별 PR 이관)
  - 비활성 (disabled profile): me5_ko (HF 401 → 폐기)

- tests/fixtures/: G0 fixture 3건 박제
  - ollama_bge_m3_embedding_response.json (G0-2: dim 1024, flat dict shape)
  - tei_embedding_response.json (G0-1: me5_large_inst, dim 1024, nested array)
  - tei_embedding_snowflake_l_v2_response.json (G0-1: snowflake, dim 1024, nested array)

운영 변경 0 (profile 격리, default up 시 미기동). production 9 컨테이너 영향 없음.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
hyungi
2026-05-23 05:04:21 +00:00
parent e4cfd81e15
commit 943ac5f59c
4 changed files with 188 additions and 0 deletions
+133
View File
@@ -0,0 +1,133 @@
# Phase 2A — Embedding candidate compose override (Diagnose only)
#
# Profile-isolated: `--profile embed-cand` 명시 opt-in. default up 시 미기동.
# production fastapi/postgres/reranker 에 영향 0.
# 본 PR 종료 시 별 chore (PR-2A-Chunks-Cand-Cleanup-1) 에서 제거.
#
# 후보 상태 (2026-05-23):
# - me5_large_inst : ✅ smoke PASS (dim 1024)
# - bge_mgemma2 : ❌ Phase 2A-Extended 별 PR 이관 (9B FP16 → VRAM OOM risk + 다운로드 cost)
# - me5_ko : ❌ 폐기 (401 Unauthorized, gated/모델명 부정확)
# - snowflake_l_v2 : 신규 추가 (Snowflake/snowflake-arctic-embed-l-v2.0, 2024-12, multilingual 강화)
#
# 사용:
# docker compose -f docker-compose.yml -f docker-compose.override.cand.yml \
# --profile embed-cand up -d embedding-cand-me5-inst
#
# 호출 (DS network 내부):
# http://embedding-cand-me5-inst:80/embed
# http://embedding-cand-snowflake-l-v2:80/embed
services:
embedding-cand-me5-inst:
image: ghcr.io/huggingface/text-embeddings-inference:1.7
container_name: hyungi_document_server-embedding-cand-me5-inst-1
expose:
- "80"
environment:
- MODEL_ID=intfloat/multilingual-e5-large-instruct
- MAX_BATCH_TOKENS=8192
- MAX_CONCURRENT_REQUESTS=4
volumes:
- embedding_cand_me5_inst_cache:/data
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: 1
capabilities: [gpu]
healthcheck:
test: ["CMD", "curl", "-fsS", "http://localhost/health"]
interval: 30s
timeout: 10s
retries: 5
start_period: 60s
profiles: ["embed-cand"]
embedding-cand-snowflake-l-v2:
image: ghcr.io/huggingface/text-embeddings-inference:1.7
container_name: hyungi_document_server-embedding-cand-snowflake-l-v2-1
expose:
- "80"
environment:
- MODEL_ID=Snowflake/snowflake-arctic-embed-l-v2.0
- MAX_BATCH_TOKENS=8192
- MAX_CONCURRENT_REQUESTS=4
volumes:
- embedding_cand_snowflake_l_v2_cache:/data
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: 1
capabilities: [gpu]
healthcheck:
test: ["CMD", "curl", "-fsS", "http://localhost/health"]
interval: 30s
timeout: 10s
retries: 5
start_period: 60s
profiles: ["embed-cand"]
# ===== 비활성 후보 (Phase 2A-Extended 별 PR 이관 또는 폐기) =====
# 진단 박제만 보존. 본 PR scope 외.
embedding-cand-bge-mgemma2:
image: ghcr.io/huggingface/text-embeddings-inference:1.7
container_name: hyungi_document_server-embedding-cand-bge-mgemma2-1
expose:
- "80"
environment:
- MODEL_ID=BAAI/bge-multilingual-gemma2
- MAX_BATCH_TOKENS=8192
- MAX_CONCURRENT_REQUESTS=4
volumes:
- embedding_cand_bge_mgemma2_cache:/data
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: 1
capabilities: [gpu]
healthcheck:
test: ["CMD", "curl", "-fsS", "http://localhost/health"]
interval: 30s
timeout: 10s
retries: 5
start_period: 300s
profiles: ["embed-cand-extended"] # 본 PR 미사용. extended 별 profile.
embedding-cand-me5-ko:
image: ghcr.io/huggingface/text-embeddings-inference:1.7
container_name: hyungi_document_server-embedding-cand-me5-ko-1
expose:
- "80"
environment:
- MODEL_ID=dragonkue/multilingual-e5-large-ko
- MAX_BATCH_TOKENS=8192
- MAX_CONCURRENT_REQUESTS=4
volumes:
- embedding_cand_me5_ko_cache:/data
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: 1
capabilities: [gpu]
healthcheck:
test: ["CMD", "curl", "-fsS", "http://localhost/health"]
interval: 30s
timeout: 10s
retries: 5
start_period: 60s
profiles: ["embed-cand-disabled"] # 401 fail. 사용 X.
volumes:
embedding_cand_me5_inst_cache:
embedding_cand_snowflake_l_v2_cache:
embedding_cand_bge_mgemma2_cache:
embedding_cand_me5_ko_cache:
+17
View File
@@ -0,0 +1,17 @@
{
"model_called": "bge-m3",
"endpoint": "http://ollama-gpu:11434/api/embeddings",
"input": "테스트 한국어 문장 + English mixed",
"response_shape": {
"embedding_dim": 1024,
"top_3_values": [
-0.010903210379183292,
-0.1681741178035736,
-0.24618254601955414
]
},
"raw_keys": [
"embedding"
],
"captured_at": "2026-05-23"
}
+19
View File
@@ -0,0 +1,19 @@
{
"slug": "me5_large_inst",
"model_id": "intfloat/multilingual-e5-large-instruct",
"endpoint": "http://embedding-cand-me5-inst:80/embed",
"input": "테스트 한국어 문장 + English mixed",
"response_shape": {
"outer_type": "list",
"batch_size": 1,
"inner_type": "list",
"embedding_dim": 1024,
"top_3_values": [
0.020381121,
-0.0070679397,
0.001018147
]
},
"note": "TEI returns nested array [[float, ...]] — Ollama 와 shape diff (Ollama = {\"embedding\": [float, ...]}). dispatcher 가 shape diff 흡수 필요.",
"captured_at": "2026-05-23"
}
@@ -0,0 +1,19 @@
{
"slug": "snowflake_l_v2",
"model_id": "Snowflake/snowflake-arctic-embed-l-v2.0",
"endpoint": "http://embedding-cand-snowflake-l-v2:80/embed",
"input": "테스트 한국어 문장 + English mixed",
"response_shape": {
"outer_type": "list",
"batch_size": 1,
"inner_type": "list",
"embedding_dim": 1024,
"top_3_values": [
0.07000499,
0.0054518348,
0.027516967
]
},
"note": "TEI nested array shape — me5_large_inst 와 동일.",
"captured_at": "2026-05-23"
}